中文版

新特性

1. GRPO：
   a. 支持GSPO算法，在GRPO训练中使用参数--importance_sampling_level sequence，参考文档：https://swift.readthedocs.io/zh-cn/latest/Instruction/GRPO/AdvancedResearch/GSPO.html
   b. GRPO server mode 支持多机 rollout，支持传入多个 vllm_server_host/port，参考脚本：https://github.com/modelscope/ms-swift/blob/main/examples/train/grpo/multi_node/server_multi_node.sh
   c. GRPO rollout 兼容 GYM 环境规范（感谢开发者Mouse的贡献），参考文档 https://swift.readthedocs.io/zh-cn/latest/Instruction/GRPO/DeveloperGuide/GYM%E7%8E%AF%E5%A2%83%E8%AE%AD%E7%BB%83.html
   d. GRPO 支持 entropy_mask 来过滤低熵token损失计算，同时logger支持记录熵值动态，参考文档https://swift.readthedocs.io/zh-cn/latest/Instruction/GRPO/AdvancedResearch/entropy_mask.html
   e. 支持多轮算法DeepEyes训练，文档参考：https://swift.readthedocs.io/zh-cn/latest/Instruction/GRPO/AdvancedResearch/deepeyes.html
   f. GRPO 支持--truncation_strategy delete，删除输入长度超过max_length的数据，并重新采样。
2. Megatron-SWIFT：
   a. 支持使LoRA训练，现支持CPT/SFT/DPO，显著加速MoE训练速度。
   - 文档参考：https://swift.readthedocs.io/zh-cn/latest/Instruction/Megatron-SWIFT%E8%AE%AD%E7%BB%83.html#lora
   - 训练脚本：https://github.com/modelscope/ms-swift/tree/main/examples/train/megatron/lora
   b. 支持loss scale，方便Agent训练，训练脚本参考：https://github.com/modelscope/ms-swift/blob/main/examples/train/megatron/lora/loss_scale.sh
   c. 默认megatron-core版本升级至0.13。
   d. 支持bshd格式，方便自定义attention_mask。
   e. 日志优化：新增GPU占用、剩余训练时间等信息打印，并输出logging.jsonl存储训练日志。
   f. 模型加载与转换速度优化，并增加模型加载进度条。
3. 训练：
   a. 支持Flash-Attention-3（含Megatron-SWIFT），训练脚本参考：https://github.com/modelscope/ms-swift/tree/main/examples/train/flash_attention_3
   b. 新增--new_speical_tokens参数，方便新增特殊tokens。训练脚本参考: https://github.com/modelscope/ms-swift/tree/main/examples/train/new_special_tokens
   c. 新增--cached_dataset参数，支持CPT/SFT的离线tokenize。训练脚本参考：https://github.com/modelscope/ms-swift/tree/main/examples/export/cached_dataset
   d. 序列Packing模块重构。加速Packing速度，并对多模态packing的磁盘存储问题优化。
   e. 支持Qwen2.5-VL混合模态数据（即单条数据中含多种模态） + deepspeed训练。
   f. 多模态模型训练支持 loss_scale。
   g. rope_scaling 支持传入字典，此外支持设置 max_model_len 对 rope_scaling 的 factor 自动调整。
   h. 支持DeepSpeed-AutoTP（该技术不支持LoRA）。
   i. 多模态Packing兼容 transformers>=4.53；序列并行兼容 transformers>=4.52。
   j. resume_only_model默认将进行数据跳过，并使用ignore_data_skip参数进行控制。
   k. MoE模型训练支持 router_aux_loss_coef 参数。
   l. template新增max_length裁剪保护机制，不对图像/视频等tokens进行裁剪。
   m. tuner_backend unsloth 支持moe模型、device_map和DDP。
   n. embedding训练支持liger_kernel。
4. RLHF：
   a. 支持MPO训练，训练脚本参考：https://github.com/modelscope/ms-swift/blob/main/examples/train/rlhf/mpo.sh
   b. 多模态DPO支持了拒绝图片输入，在数据集中加入rejected_images列。
5. 推理部署：
   a. 支持embedding系列模型的推理部署，包括pt/vllm/sglang的infer_backend。部署脚本参考：https://github.com/modelscope/ms-swift/tree/main/examples/deploy/embedding
   b. InferEngine支持return_details参数，以输出prompt_token_ids和token_ids。
   c. vLLM推理引擎兼容更多多模态模型：ovis2, glm4_1v, keye-vl, kimi-vl, glm4v, phi4-multimodal, llama4。
   d. vLLM参数重构，参数名前加入vllm_前缀。GRPO模块复用vLLM参数。
6. 导出：
   a. QLoRA支持Merge-LoRA，脚本参考：https://github.com/modelscope/ms-swift/tree/main/examples/train/qlora
   b. 支持MoE/多模态模型的FP8/BNB量化，脚本参考：https://github.com/modelscope/ms-swift/tree/main/examples/export/quantize

新模型

1. 纯文本模型：
   a. Qwen/Qwen3-235B-A22B-[Instruct/Thinking]-2507, Qwen/Qwen3-Coder-480B-A35B-Instruct, Qwen/Qwen3-4B-[Instruct/Thinking]-2507系列（含Megatron-SWIFT），训练脚本参考：#5033
   b. openai-mirror/gpt-oss-20b系列，最佳实践参考：#5277
   c. ZhipuAI/GLM-4.5系列（含Megatron-SWIFT），训练脚本参考：https://github.com/modelscope/ms-swift/blob/main/examples/train/megatron/lora/glm4_5_106b.sh
   d. Hunyuan-7B-Instruct系列，最佳实践参考：#5236
   e. mistralai/Devstral-Small-2505
2. 多模态模型：
   a. OpenBMB/MiniCPM-V-4，训练脚本参考：https://github.com/modelscope/ms-swift/blob/main/examples/models/minicpmv/train.sh

English Version

New Features

1. GRPO
   a. Added support for the GSPO algorithm. Use --importance_sampling_level sequence during GRPO training. Docs: https://swift.readthedocs.io/en/latest/Instruction/GRPO/AdvancedResearch/GSPO.html
   b. GRPO “server mode” now supports multi-node rollout; pass in multiple vllm_server_host/port. Example script: https://github.com/modelscope/ms-swift/blob/main/examples/train/grpo/multi_node/server_multi_node.sh
   c. GRPO rollout is now GYM-compatible (thanks to contributor Mouse). Docs: https://swift.readthedocs.io/en/latest/Instruction/GRPO/DeveloperGuide/gym_env.html
   d. Added entropy_mask for filtering low-entropy tokens during loss computation, and the logger now tracks entropy dynamics. Docs: https://swift.readthedocs.io/en/latest/Instruction/GRPO/AdvancedResearch/entropy_mask.html
   e. Added support for the multi-round DeepEyes algorithm. Docs: https://swift.readthedocs.io/en/latest/Instruction/GRPO/AdvancedResearch/deepeyes.html
   f. GRPO supports --truncation_strategy delete: remove samples whose input length exceeds max_length and resample.
2. Megatron-SWIFT
   a. Added LoRA training (CPT/SFT/DPO) to significantly accelerate MoE training.
   - Docs: https://swift.readthedocs.io/en/latest/Instruction/Megatron-SWIFT-Training.html#lora-training
   - Scripts: https://github.com/modelscope/ms-swift/tree/main/examples/train/megatron/lora
   b. Added loss-scaling to simplify Agent training. Script: https://github.com/modelscope/ms-swift/blob/main/examples/train/megatron/lora/loss_scale.sh
   c. Default megatron-core upgraded to 0.13.
   d. Added bshd tensor format to facilitate custom attention_mask.
   e. Logging improvements: prints GPU memory, estimated remaining time, and writes logging.jsonl.
   f. Faster model loading & conversion plus a progress bar.
3. Training
   a. Added Flash-Attention-3 support (including Megatron-SWIFT). Scripts: https://github.com/modelscope/ms-swift/tree/main/examples/train/flash_attention_3
   b. New --new_special_tokens flag for adding special tokens. Scripts: https://github.com/modelscope/ms-swift/tree/main/examples/train/new_special_tokens
   c. New --cached_dataset flag for offline tokenization in CPT/SFT. Scripts: https://github.com/modelscope/ms-swift/tree/main/examples/export/cached_dataset
   d. Re-implemented the sequence-packing module for faster packing and better multimodal disk I/O.
   e. Qwen2.5-VL hybrid-modal data (multiple modalities in a single sample) + DeepSpeed training supported.
   f. Multimodal training now supports loss-scaling.
   g. rope_scaling now accepts a dict; max_model_len can auto-adjust the scaling factor.
   h. Added DeepSpeed-AutoTP (not compatible with LoRA).
   i. Multimodal packing is compatible with transformers ≥ 4.53; sequence parallelism with transformers ≥ 4.52.
   j. With resume_only_model, data skipping is enabled by default; control via ignore_data_skip.
   k. MoE training supports router_aux_loss_coef.
   l. Template files get a max_length clipping safeguard (no clipping of image/video tokens).
   m. tuner_backend unsloth now supports MoE models, device_map, and DDP.
   n. Embedding training supports liger_kernel.
4. RLHF
   a. Added MPO training. Script: https://github.com/modelscope/ms-swift/blob/main/examples/train/rlhf/mpo.sh
   b. Multimodal DPO can now reject image inputs by adding a rejected_images column.
5. Inference & Deployment
   a. Added deployment for embedding models across pt/vllm/sglang back-ends. Scripts: https://github.com/modelscope/ms-swift/tree/main/examples/deploy/embedding
   b. InferEngine supports return_details to output prompt_token_ids and token_ids.
   c. vLLM back-end now supports more multimodal models: ovis2, glm4_1v, keye-vl, kimi-vl, glm4v, phi4-multimodal, llama4.
   d. vLLM arguments refactored: all start with the vllm_ prefix. GRPO module reuses the same options.
6. Export
   a. QLoRA now supports Merge-LoRA. Scripts: https://github.com/modelscope/ms-swift/tree/main/examples/train/qlora
   b. Added FP8 / BNB quantization for MoE and multimodal models. Scripts: https://github.com/modelscope/ms-swift/tree/main/examples/export/quantize

New Models

1. Text-only
   a. Qwen/Qwen3-235B-A22B-[Instruct/Thinking]-2507, Qwen/Qwen3-Coder-480B-A35B-Instruct, and Qwen/Qwen3-4B-[Instruct/Thinking]-2507 (Megatron-SWIFT supported). Training script: #5033
   b. openai-mirror/gpt-oss-20b family. Best-practice: #5277
   c. ZhipuAI/GLM-4.5 family (Megatron-SWIFT supported). Training script: https://github.com/modelscope/ms-swift/blob/main/examples/train/megatron/lora/glm4_5_106b.sh
   d. Hunyuan-7B-Instruct family. Best-practice: #5236
   e. mistralai/Devstral-Small-2505
2. Multimodal
   a. OpenBMB/MiniCPM-V-4. Training script: https://github.com/modelscope/ms-swift/blob/main/examples/models/minicpmv/train.sh

What's Changed

• [grpo] fix server arg check by @hjh0119 in #4865
• [SP] clean up imports by @hjh0119 in #4878
• fix loss_scale sp by @tastelikefeet in #4880
• fix seq_cls generation_config by @Jintao-Huang in #4882
• optimize imports by @tastelikefeet in #4883
• [model] fix qwen eos_token by @Jintao-Huang in #4888
• Fix: Correct training hang for Keye-VL on DeepSpeed with mixed data by @0russwest0 in #4889
• [megatron] support LoRA & support loss_scale by @Jintao-Huang in #4812
• update framework.txt by @Jintao-Huang in #4896
• [megatron] fix pp mla by @Jintao-Huang in https://gi...

Full Changelog: v3.6.3...v3.6.4

Full Changelog: v3.6.2...v3.6.3

Full Changelog: v3.6.1...v3.6.2

Full Changelog: v3.6.0...v3.6.1

中文版

新特性

1. Megatron-SWIFT：
   a. 支持更多的 MoE 模型结构，包括：DeepseekV3ForCausalLM、Dots1ForCausalLM 和 Ernie4_5_MoeForCausalLM。训练脚本参考：https://github.com/modelscope/ms-swift/tree/main/examples/train/megatron/moe
   b. 支持更多的 Dense 模型结构，包括：MiMoForCausalLM、InternLM3ForCausalLM 和 Ernie4_5_ForCausalLM。训练脚本参考：https://github.com/modelscope/ms-swift/tree/main/examples/train/megatron/dense
   c. 支持 DPO 训练。训练脚本参考：https://github.com/modelscope/ms-swift/tree/main/examples/train/megatron/rlhf/dpo
   d. 支持 FP8 训练。
   e. 支持更多 rope scaling 类型，包括：default、linear、yarn、dynamic、longrope、llama3 等。
   f. --test_convert_precision参数优化，方便测试 mcore 与 huggingface 模型权重转换精度。
2. GRPO：
   a. GRPO 多轮训练重构，支持使用 AsyncEngine 加速多轮推理，参考文档：https://swift.readthedocs.io/zh-cn/latest/Instruction/GRPO/DeveloperGuide/%E5%A4%9A%E8%BD%AE%E8%AE%AD%E7%BB%83.html
   b. offload_model 参数额外对参考模型进行卸载。
   c. 优化 sleep_level 和 offload_model 参数下的显存管理。
   d. reward_funcs 增加了 trainer_state 入参，方便获取当前训练步数和总步数。
3. 训练：
   a. 支持 reranker 训练，训练脚本参考：https://github.com/modelscope/ms-swift/tree/main/examples/train/reranker
   b. CPT/SFT/DPO/GRPO 纯文本大模型训练支持 ring-attention 切分序列长度，降低显存占用。训练脚本参考：https://github.com/modelscope/ms-swift/tree/main/examples/train/long_text/ring_attention
   c. channel loss 在CPT/SFT训练时，兼容 padding_free 与 packing。 感谢招商银行技术团队的贡献。
   d. remove_unused_columns 参数优化。设置为 False，则将额外数据集传递至 Trainer 内，方便自定义损失函数。
   e. split_dataset_ratio参数默认值从0.01修改为0，默认不再进行验证集切分，需要手动设置--split_dataset_ratio或者--val_dataset。
   f. 多模态模型 packing/padding_free 损失对齐问题修复。详见此PR：#4838
   g. swanlab 支持训练完成后的飞书通知回调。
4. RLHF：
   a. 纯文本/多模态模型支持 GKD 训练，部分场景下支持 padding_free 和 packing，训练脚本如下：
   i. 大模型：https://github.com/modelscope/ms-swift/blob/main/examples/train/rlhf/gkd.sh
   ii. 多模态大模型：https://github.com/modelscope/ms-swift/blob/main/examples/train/multimodal/rlhf/gkd.sh
   b. reward model 训练支持 margin 参数支持，参考文档：https://swift.readthedocs.io/zh-cn/latest/Instruction/%E4%BA%BA%E7%B1%BB%E5%AF%B9%E9%BD%90.html#rm
5. 全链路：
   a. 支持使用 SGLang 推理引擎对 ms-swift 推理/部署/评测/ui模块进行加速，设置--infer_backend sglang即可。推理脚本参考：https://github.com/modelscope/ms-swift/tree/main/examples/infer/sglang
   b. 支持 FP8 量化，量化脚本参考：https://github.com/modelscope/ms-swift/blob/main/examples/export/quantize/fp8.sh
6. Web-UI：
   a. 支持 SFT/RLHF/GRPO 在不同 Tab 页面训练，支持保存训练命令行。
   b. Web-UI 界面支持数据采样。

新模型

1. 多模态模型：
   a. ZhipuAI/GLM-4.1V-9B-Thinking系列
   b. Kwai-Keye/Keye-VL-8B-Preview
   c. moonshotai/Kimi-VL-A3B-Thinking-2506
   d. google/gemma-3n-E2B-it系列
2. 纯文本模型：
   a. PaddlePaddle/ERNIE-4.5-21B-A3B-PT系列
   b. rednote-hilab/dots.llm1.inst系列
   c. Tencent-Hunyuan/Hunyuan-A13B-Instruct
   d. MiniMax/MiniMax-M1-80k系列（推理）
   e. moonshotai/Kimi-Dev-72B
   f. cognitivecomputations/DeepSeek-R1-0528-AWQ

English Version

New Features

1. Megatron-SWIFT:
   a. Support for more MoE model architectures, including: DeepseekV3ForCausalLM, Dots1ForCausalLM, and Ernie4_5_MoeForCausalLM. Training script reference: https://github.com/modelscope/ms-swift/tree/main/examples/train/megatron/moe
   b. Support for more Dense model architectures, including: MiMoForCausalLM, InternLM3ForCausalLM, and Ernie4_5_ForCausalLM. Training script reference: https://github.com/modelscope/ms-swift/tree/main/examples/train/megatron/dense
   c. DPO training supported. Training script reference: https://github.com/modelscope/ms-swift/tree/main/examples/train/megatron/rlhf/dpo
   d. FP8 training supported.
   e. More rope scaling types supported, including: default, linear, yarn, dynamic, longrope, llama3, etc.
   f. --test_convert_precision parameter optimized for easier testing of weight conversion precision between mcore and huggingface models.
2. GRPO:
   a. GRPO multi-turn training refactored, supporting accelerated multi-turn inference with AsyncEngine. Documentation: https://swift.readthedocs.io/zh-cn/latest/Instruction/GRPO/DeveloperGuide/%E5%A4%9A%E8%BD%AE%E8%AE%AD%E7%BB%83.html
   b. The offload_model parameter now also offloads the reference model.
   c. Optimized GPU memory management under sleep_level and offload_model parameters.
   d. Added trainer_state as an input parameter to reward_funcs, making it easier to obtain the current and total training steps.
3. Training:
   a. Reranker training supported. Training script reference: https://github.com/modelscope/ms-swift/tree/main/examples/train/reranker
   b. CPT/SFT/DPO/GRPO pure-text large model training supports ring-attention sequence length partitioning, reducing memory usage. Training script reference: https://github.com/modelscope/ms-swift/tree/main/examples/train/long_text/ring_attention
   c. Channel loss in CPT/SFT training is compatible with padding_free and packing. Thanks to the technical team at China Merchants Bank for their contribution.
   d. Optimized remove_unused_columns parameter. When set to False, extra dataset columns are passed to the Trainer for custom loss functions.
   e. The default value for split_dataset_ratio changed from 0.01 to 0, so the validation set is not split by default. You now need to manually set --split_dataset_ratio or --val_dataset.
   f. Fixed loss alignment issue between packing/padding_free for multimodal models. For details, see this PR: #4838
   g. Swanlab now supports Feishu (Lark Suite) notification callback after training is completed.
4. RLHF:
   a. Pure-text and multimodal models support GKD training, with some scenarios supporting padding_free and packing. Training scripts:
   i. Large models: https://github.com/modelscope/ms-swift/blob/main/examples/train/rlhf/gkd.sh
   ii. Multimodal large models: https://github.com/modelscope/ms-swift/blob/main/examples/train/multimodal/rlhf/gkd.sh
   b. Reward model training now supports the margin parameter. Documentation: https://swift.readthedocs.io/zh-cn/latest/Instruction/%E4%BA%BA%E7%B1%BB%E5%AF%B9%E9%BD%90.html#rm
5. Full Pipeline:
   a. SGLang inference engine can be used to accelerate ms-swift inference/deployment/evaluation/ui modules, by setting --infer_backend sglang. Inference script reference: https://github.com/modelscope/ms-swift/tree/main/examples/infer/sglang
   b. FP8 quantization supported. Quantization script reference: https://github.com/modelscope/ms-swift/blob/main/examples/export/quantize/fp8.sh
6. Web-UI:
   a. Supports SFT/RLHF/GRPO training on different Tab pages, and saves training command lines.
   b. Web-UI interface supports data sampling.

New Models

1. Multimodal Models:
   a. ZhipuAI/GLM-4.1V-9B-Thinking series
   b. Kwai-Keye/Keye-VL-8B-Preview
   c. moonshotai/Kimi-VL-A3B-Thinking-2506
   d. google/gemma-3n-E2B-it series
2. Pure Text Models:
   a. PaddlePaddle/ERNIE-4.5-21B-A3B-PT series
   b. rednote-hilab/dots.llm1.inst series
   c. Tencent-Hunyuan/Hunyuan-A13B-Instruct
   d. MiniMax/MiniMax-M1-80k series (inference)
   e. moonshotai/Kimi-Dev-72B
   f. cognitivecomputations/DeepSeek-R1-0528-AWQ

What's Changed

• fix emb script and docs by @tastelikefeet in #4521
• [grpo] update doc about move_model_batches by @hjh0119 in #4523
• fix LoraModel by @Jintao-Huang in #4536
• support cognitivecomputations/DeepSeek-R1-0528-AWQ by @Jintao-Huang in #4537
• fix: handle INFONCE_HARD_NEGATIVES as integer if provided by @dlutwy in #4545
• fix qwen3 embedding saving by @tastelikefeet in #4548
• [megatron/dpo] fix megatron packing_cache & update DPOTrainer by @Jintao-Huang in #4556
• [megatron] support DPO by @Jintao-Huang in #4193
• support dots1 by @Jintao-Huang in #4560
• [grpo] support offloading reference model by @hjh0119 in #4554
• [grpo] fix the pickle data collator by @hjh0119 in #4562
• [dataset] fix toolbench (local) by @Jintao-Huang in #4563
• [Bug]Fix ulysses train steps, embedding negative sample length by @tastelikefeet in #4565
• fix args.json by @Jintao-Huang in #4566
• [model] fix ovis gradient_checkpointing vit no_grad by @Jintao-Huang in #4571
• [megatron] Fix megatron all_reduce warning by @Jintao-Huang in #4568
• [grpo] remove data collator to top-level to avoid pickle error in spawn mode by @hjh0119 in #4582
• [grpo] model weight synchronization before first turn rollout with async generation by @hjh0119 in #4584
• [megatron] support more rope_scaling & support deepseek-r1-qwen3-8b/internlm3/mimo-7b by @Jintao-Huang in #4576
• [grpo] restore num_generations check by @hjh0119 in #4590
• fix gc_kwargs by @Jintao-Huang in #4591
• Fix UI llm_train by @slin000111 in #4592
• [mirror] update swift mirror by @Jintao-Huang in #4601
• [megatron] compat megatron-core main branch by @Jintao-Huang in https://github.com/modelscope/ms-swift...

Full Changelog: v3.5.2...v3.5.3

Full Changelog: v3.5.1...v3.5.2

Full Changelog: v3.5.0...v3.5.1

中文版

新特性

1. GRPO：
   a. 代码重构，使用参数vllm_mode指定。参数说明详见参考文档：https://swift.readthedocs.io/zh-cn/latest/Instruction/GRPO.html#id1:~:text=vllm_mode%20server%20%E5%8F%82%E6%95%B0,colocate%20mode%20%E7%94%9F%E6%95%88%E3%80%82
   b. GRPO长文本优化，支持ulysses序列并行，显著降低长文本训练显存占用，训练脚本参考：https://github.com/modelscope/ms-swift/blob/main/examples/train/long_text/sequence_parallel_grpo.sh
   c. 新增sync_ref_model参数，支持训练中同步参考模型权重。
   d. 支持 liger kernel loss，使用参数 use_liger_kernel，降低显存占用。
   e. External mode 支持 move_model_batches，降低zero3同步权重时的显存峰值。
   f. 集成 INTELLECT-2 的 Two-Sided Clipping 算法，使用参数 delta。
   g. 支持奖励函数返回 None，适用于多任务训练，参考文档：https://swift.readthedocs.io/zh-cn/latest/Instruction/GRPO.html#id7
   h. Internal mode 支持 vllm_server_base_url，传入外部 vLLM 服务器url。
   i. 插件拓展：支持 QwenLong-L1 奖励模型插件。
   j. 新增 steps_per_generation/generation_batch_size 参数，支持自定义采样批量大小。
   k. Web-UI支持GRPO训练。
   l. 以下参数将在 v3.6 移除：tensor_parallel_size / vllm_device / vllm_max_num_seqs / num_infer_workers。
2. 训练：
   a. CPT/SFT/DPO/GRPO 支持 padding free。通过将批次数据展平避免数据填充（padding），显著降低显存并加速训练。训练脚本参考：https://github.com/modelscope/ms-swift/tree/main/examples/train/padding_free
   b. 多模态训练增强。支持使用 vit_lr 和 aligner_lr 参数独立控制 ViT 和 Aligner 模块的学习率。支持通过 vit_gradient_checkpointing 参数单独控制 vit 模块的 gradient checkpointing，性能基准测试参考：https://github.com/modelscope/ms-swift/blob/main/examples/train/multimodal/vit_gradient_checkpointing.sh
   c. CPT/SFT支持使用 channel loss 对不同 channel 数据集分别统计损失值。感谢招商银行技术团队的贡献。
   d. CPT/SFT/DPO支持 use_logits_to_keep参数，降低显存占用，提升训练速度。
   e. Qwen2.5-VL/Omni 支持传入图像目录进行视频训练。
3. 推理部署：
   a. swift infer批处理优化，新增 write_batch_size 参数，用于控制批处理推理结果写入result_path的间隔。
   b. vllm 推理引擎默认使用 V1 engine，并支持TP和DP结合的推理模式，脚本参考：https://github.com/modelscope/ms-swift/blob/main/examples/infer/vllm/dp_tp.sh
4. Megatron-SWIFT：
   a. 非流式数据集支持通过 max_epochs 自动计算 train_iters。
   b. 提供 extra_megatron_kwargs 参数，支持未写入ms-swift的megatron参数传入。

新模型

1. Qwen/Qwen3-Embedding-0.6B系列，训练脚本参考：https://github.com/modelscope/ms-swift/blob/main/examples/train/embedding/train_emb.sh
2. deepseek-ai/DeepSeek-R1-0528-Qwen3-8B系列，最佳实践参考https://mp.weixin.qq.com/s/-hhfGiiGTqXUybwPH525gw
3. iic/QwenLong-L1-32B
4. XiaomiMiMo/MiMo-7B-RL-0530、XiaomiMiMo/MiMo-VL-7B-SFT系列
5. OpenBMB/MiniCPM4-0.5B系列

English Version

New Features

1. GRPO:
   a. Code refactored, specified via the vllm_mode parameter. For details, refer to the documentation: https://swift.readthedocs.io/en/latest/Instruction/GRPO.html#arguments-and-execution-script:~:text=vllm_mode%20server%20parameter,in%20colocate%20mode.
   b. GRPO long-text optimization with Ulysses sequence parallelism, significantly reducing GPU memory usage during long-text training. Training script: https://github.com/modelscope/ms-swift/blob/main/examples/train/long_text/sequence_parallel_grpo.sh
   c. Added sync_ref_model parameter to synchronize reference model weights during training.
   d. Supports Liger Kernel Loss via use_liger_kernel parameter, reducing GPU memory consumption.
   e. External mode supports move_model_batches to lower peak GPU memory during ZeRO-3 weight synchronization.
   f. Integrated INTELLECT-2’s Two-Sided Clipping algorithm using the delta parameter.
   g. Supports reward functions returning None, applicable for multi-task training. For details, refer to the documentation: https://swift.readthedocs.io/en/latest/Instruction/GRPO.html#multi-task-training
   h. Internal mode supports vllm_server_base_url for passing external vLLM server URLs.
   i. Plugin extension: Added QwenLong-L1 reward model plugin.
   j. Added steps_per_generation and generation_batch_size parameters for customizing sampling batch size.
   k. Web-UI supports GRPO training.
   l. The following parameters will be deprecated in v3.6: tensor_parallel_size, vllm_device, vllm_max_num_seqs, num_infer_workers.
2. Training:
   a. CPT/SFT/DPO/GRPO support padding-free training. By flattening batch data to avoid padding, GPU memory usage is reduced and training speed is improved. Script: https://github.com/modelscope/ms-swift/tree/main/examples/train/padding_free
   b. Multimodal training enhancements: Supports separate learning rates for ViT and Aligner modules via vit_lr and aligner_lr parameters. Added vit_gradient_checkpointing to independently control gradient checkpointing for ViT modules. Benchmark: https://github.com/modelscope/ms-swift/blob/main/examples/train/multimodal/vit_gradient_checkpointing.sh
   c. CPT/SFT support channel_loss to separately calculate loss for different channel datasets. Thanks to the contributions from the technical team at China Merchants Bank.
   d. CPT/SFT/DPO support use_logits_to_keep to reduce GPU memory usage and accelerate training.
   e. Qwen2.5-VL/Omni support video training by passing image directories.
3. Inference & Deployment:
   a. Optimized swift infer batching with new write_batch_size parameter to control inference result write intervals to result_path.
   b. vLLM inference engine now defaults to V1 engine and supports hybrid Tensor Parallelism (TP) and Data Parallelism (DP). Script: https://github.com/modelscope/ms-swift/blob/main/examples/infer/vllm/dp_tp.sh
4. Megatron-SWIFT:
   a. Non-streaming datasets automatically calculate train_iters via max_epochs.
   b. Added extra_megatron_kwargs to pass unlisted Megatron parameters into ms-swift.

New Models

1. Qwen/Qwen3-Embedding-0.6B series. Training script reference: https://github.com/modelscope/ms-swift/blob/main/examples/train/embedding/train_emb.sh
2. deepseek-ai/DeepSeek-R1-0528-Qwen3-8B series. Best practices: https://mp.weixin.qq.com/s/-hhfGiiGTqXUybwPH525gw
3. iic/QwenLong-L1-32B
4. XiaomiMiMo/MiMo-7B-RL-0530 & XiaomiMiMo/MiMo-VL-7B-SFT series
5. OpenBMB/MiniCPM4-0.5B series

What's Changed

• [grpo] code refactor by @hjh0119 in #4097
• support yarn by @tastelikefeet in #4197
• fix ppo init model by @hjh0119 in #4199
• fix ppo reward model by @hjh0119 in #4200
• [doc] remove vllm version warning in grpo by @hjh0119 in #4204
• [grpo] fix colocate + tp by @hjh0119 in #4209
• Refactor packing by @Jintao-Huang in #4207
• [grpo] set system in inputs by @hjh0119 in #4214
• fix mm packing by @Jintao-Huang in #4217
• fix packing multi_node by @Jintao-Huang in #4222
• fix get reward model by @hjh0119 in #4225
• fix val_dataset_shuffle by @Jintao-Huang in #4226
• fix task type judgement in rlhf by @hjh0119 in #4228
• fix eval extral args by @Yunnglin in #4227
• fix loss_scale by @Jintao-Huang in #4229
• update docs by @Jintao-Huang in #4235
• [rlhf] prepare_model for ref_model & reduce peak memory in dpo by @hjh0119 in #4232
• fix qwen2_5_vl VIDEO_TOTAL_PIXELS by @Jintao-Huang in #4236
• Support super long length sft by @tastelikefeet in #4237
• compat transformers 4.52 by @Jintao-Huang in #4238
• update liger_kernel docs by @Jintao-Huang in #4241
• [grpo] support synchronizing ref model by @hjh0119 in #4242
• optimize packing io by @Jintao-Huang in #4244
• fix register_post_encode_hook by @Jintao-Huang in #4247
• compat megatron-core 0.11 by @Jintao-Huang in #4250
• fix qwen2_5_omni by @Jintao-Huang in #4253
• fix readme by @Jintao-Huang in #4256
• [grpo] set v1 engine as default in external rollout by @hjh0119 in #4258
• fix ddp_timeout by @Jintao-Huang in #4259
• Add tqdm by @Jintao-Huang in #4260
• Fix is_master by @Jintao-Huang in #4262
• fix ppo zero3 by @Jintao-Huang in #4263
• test link valid by @Jintao-Huang in #4265
• update docs & fix quant by @Jintao-Huang in #4268
• [grpo] fix external mode&multi turn by @hjh0119 in #4255
• fix ulysses eval by @tastelikefeet in #4271
• support IndexedDataset shard by @Jintao-Huang in #4269
• Support vit_lr aligner_lr by @Jintao-Huang in #4273
• support padding_free CPT/SFT by @Jintao-Huang in #4274
• [grpo] fix num of reward_model > 1 by @hjh0119 in #4287
• fix n > 1 with vLLM V1 Engine by @hjh0119 in #4295
• update load_args by @Jintao-Huang in #4296
• update swift image by @Jintao-Huang in #4309
• Fix ulysses pending by @tastelikefeet in https://github...